View
8
Download
0
Category
Preview:
Citation preview
arX
iv:2
105.
0962
8v1
[cs
.IT
] 2
0 M
ay 2
021
1
Deterministic Pilot Design and Channel Estimation
for Downlink Massive MIMO-OTFS Systems in
Presence of the Fractional DopplerDing Shi, Graduate Student Member, IEEE, Wenjin Wang, Member, IEEE, Li You, Member, IEEE,
Xiaohang Song, Member, IEEE, Yi Hong, Senior Member, IEEE, Xiqi Gao, Fellow, IEEE,
and Gerhard Fettweis, Fellow, IEEE
Abstract—Although the combination of the orthogonal timefrequency space (OTFS) modulation and the massive multiple-input multiple-output (MIMO) technology can make communica-tion systems perform better in high-mobility scenarios, there arestill many challenges in downlink channel estimation owing toinaccurate modeling and high pilot overhead in practical systems.In this paper, we propose a channel state information (CSI) ac-quisition scheme for downlink massive MIMO-OTFS in presenceof the fractional Doppler, including deterministic pilot designand channel estimation algorithm. First, we analyze the input-output relationship of the single-input single-output (SISO) OTFSbased on the orthogonal frequency division multiplexing (OFDM)modem and extend it to massive MIMO-OTFS. Moreover, weformulate an accurate model for the practical system in whichthe fractional Doppler is considered and the influence of subpathsis revealed. A deterministic pilot design is then proposed basedon the model and the structure of the pilot matrix to reduce pilotoverhead and save memory consumption. Since channel geometrychanges very slowly relative to the communication timescale, weput forward a modified sensing matrix based channel estimation(MSMCE) algorithm to acquire the downlink CSI. Simulationresults demonstrate that the proposed downlink CSI acquisitionscheme has significant advantages over traditional algorithms.
Index Terms—OTFS, massive MIMO, fractional Doppler,channel estimation, deterministic pilot.
I. INTRODUCTION
TO meet the needs of future wireless communications,
especially various new services and scenarios that are
continually emerging, the fifth-generation (5G) mobile systems
come into being. One objective of 5G is to achieve reliable
communication for the scenarios with high Doppler spread,
such as high-mobility scenarios [1], [2]. Orthogonal frequency
This work was supported in part by the National Key Research andDevelopment Program of China under Grant 2018YFB1801103; in part by theJiangsu Province Basic Research Project under Grant BK20192002; in part bythe National Natural Science Foundation of China under Grants 61761136016,61631018, and 61801114; in part by the Fundamental Research Funds forthe Central Universities; and in part by the German Science Foundation(DFG) through the Project Large-Scale and Hierarchical Bayesian Inferencefor Future Mobile Communication Networks under Grant 392016367.
D. Shi, W. J. Wang, L. You and X. Q. Gao are with the National MobileCommunications Research Laboratory, Southeast University, Nanjing 210096,China (e-mail: shiding@seu.edu.cn; wangwj@seu.edu.cn; liyou@seu.edu.cn;xqgao@seu.edu.cn).
X. H. Song and G. Fettweis are with the Vodafone Chair Mobile Commu-nications Systems, Technische Universitat Dresden, 01062 Dresden, Germany(e-mail: xiaohang.song@tu-dresden.de; gerhard.fettweis@tu-dresden.de).
Y. Hong is with the ECSE Department, Monash University, Clayton, VIC3800, Australia (e-mail: Yi.Hong@monash.edu).
division multiplexing (OFDM) technology has been widely
used to combat the inter-symbol interference (ISI) in time-
invariant channels [3]–[6]. However, when there is a high
Doppler spread in time-variant channels, which would lead
to severe inter-carrier interference (ICI), the performance of
OFDM degrades significantly [7].
To cope with this problem, orthogonal time frequency
space (OTFS) modulation technology was proposed [8]–[10]
and attracted much attention due to significant advantages
in time-variant channels. OTFS converts the signal in the
time-frequency domain into a delay-Doppler domain and per-
forms modulation accordingly. In particular, each information
symbol in the delay-Doppler domain is expressed by a pair
of orthogonal basis functions in the time-frequency domain,
i.e., full diversity in the time-frequency domain. Furthermore,
transmitted symbols experience roughly constant channels
even with high Doppler spread. Another advantage is that the
implementation of OTFS can be achieved simply as an overlay
of the OFDM systems. Farhang et al. investigated the OTFS
system based on OFDM [11] and designed a modem scheme
with low complexity, where the cyclic prefix (CP) is added in
front of each OFDM symbol in an OTFS symbol.
Channel estimation is crucial in OTFS systems. In [12] and
[13], training-based channel estimation methods were inves-
tigated. In [12], Murali et al. used the pseudo-random (PN)
sequence as pilots to acquire the channel state information
(CSI). In [13], Raviteja et al. arranged the pilots and data in the
same OTFS symbol and estimated the delay-Doppler channel
by a threshold method. After the acquisition of the CSI, the
symbol detection can be performed. In [12], the Markov chain
Monte Carlo (MCMC) sampling was utilized for the low-
complexity symbol detection. In [14], an explicit derivation of
the input-output relationship of OTFS systems was given, and
the effect of fractional Doppler caused by insufficient sampling
of Doppler-dimension was analyzed. They also developed
a message passing (MP) method to detect OTFS symbols.
However, when they used the rectangular pulse-shaping wave-
forms, the CP was not considered, which would cause ISI and
might make the receiver more complicated. Different from
the symbol spaced sampling framework in [14], Ge et al.
designed a fractionally sampling scheme for symbol detection
[15], where only one CP is inserted for the whole OTFS
symbol. Compared with [11], less CP overhead in [13]–[15]
can improve spectral efficiency. In [16], Wei et al. reduced the
2
channel spread caused by the fractional Doppler and improved
the channel estimation performance through the design of
transmitter and receiver windows. In [17], the embedded pilot-
aided CSI acquisition scheme proposed in [13] and the MP
algorithm proposed in [14] were extended to multiple-input
multiple-output (MIMO) OTFS systems.
Massive MIMO technology has been one of the key
technologies of 5G owing to the immense improvement of
spectrum and power efficiencies [18]–[26]. Considering high
Doppler spread scenarios, the integration of massive MIMO
and OTFS can further improve the performance. To utilize
such benefits, the base station (BS) requires downlink CSI
for precoding. In traditional time-division duplexing (TDD)
systems, uplink training can be used to acquire the downlink
CSI via leveraging the channel reciprocity [18], [27]. However,
when systems work in frequency-division duplexing (FDD)
mode, user terminals need the pilots transmitted by the BS
to acquire the CSI and feed it back. Thus the downlink
channel estimation is necessary for massive MIMO-OTFS
systems. In [28], Shen et al. demonstrated the 3-dimensional
sparsity of the downlink massive MIMO-OTFS channel, based
on which a 3D-structured orthogonal matching pursuit (3D-
SOMP) method was proposed to acquire the downlink CSI.
However, the pilot used for the channel estimation is the
random one, which is hard to be implemented and consumes
much memory for its randomness in the practical system.
In [29], Liu et al. perform the downlink channel estimation
with the help of channel parameters obtained from uplink
training. But downlink CSI acquisition can not be performed
independently and must be after the uplink channel estimation.
In [30], Shan et al. designed a low-overhead pilot pattern
for the CSI acquisition and equalized the received signal
from each angle through the channel information obtained.
It is noticed that the fractional Doppler is not considered in
[28]–[30], and the system models are accurate only when the
Doppler frequency of each path of the channel is mapped to an
integer tap. It is impractical since the resolution of the Doppler
axis is not sufficient. In [31], Li et al. developed a path division
multiple access (PDMA) scheme by which different users used
scheduled delay-Doppler domain grids to communicate with
BS simultaneously without inter-user interference. However,
they only considered TDD systems, where downlink CSI
can be obtained by the reciprocity between the uplink and
downlink channel. Thus only the uplink channel estimation
was discussed, and the proposed scheme is not suitable for
FDD systems.
In this work, we consider the fractional Doppler caused
by insufficient sampling of Doppler-dimension in practical
systems. Moreover, for massive MIMO-OTFS systems, we
propose a downlink CSI acquisition scheme, including deter-
ministic pilot design and channel estimation algorithm. The
contributions are summarized as follows:
• In the single-input single-output (SISO) case, we analyze
the input-output relationship of the OTFS system, where
the OFDM modem is chosen as the time-frequency
modem. Specifically, we consider that the Doppler fre-
quency of each path is mapped to an integer and a
fractional Dopper tap and reveal the influence of the
subpaths contained in each dominant path. Moreover, the
characteristics of subpaths are used to simplify such a
system model. Then we extend it to the massive MIMO-
OTFS system and establish the downlink CSI acquisition
model.
• To reconstruct the channel accurately, we design a
Zadoff-Chu (ZC) sequence based deterministic pilot,
which can achieve better sensing performance and mem-
ory consumption saving than the random pilot used in
[28]. We first analyze the relationship between the pilot
matrix and the position of pilots at each beam. Then we
give two conditions that the deterministic pilots need to
satisfy to ensure the low coherence between the pilot
matrix columns. Next, ZC sequences are chosen as pilots,
whose performance is also discussed.
• Based on the modified model, we propose a modified
sensing matrix based channel estimation (MSMCE) algo-
rithm to acquire CSI. Due to the slow change of delay-
Doppler channels, path delays and Doppler frequencies
of all dominant paths are extracted from the previous
channel estimation result and used to modify the sensing
matrix for more accurate CSI acquisition. Then we can
utilize the estimated CSI to update the path delays and
Doppler frequencies, which can be used in the next
channel estimation.
The rest of the paper is organized as follows. In Section II,
we investigate models for SISO OTFS systems and massive
MIMO-OTFS systems. Section III establish a downlink CSI
acquisition model. Based on the model, we design a determin-
istic pilot matrix and propose a MSMCE algorithm. Section
IV presents the performance of the proposed CSI acquisition
scheme by simulation results, and the paper is concluded in
Section V.
Notations: The superscripts (·)∗ and (·)H denote the con-
jugate and conjugated-transpose operations, respectively. The
uppercase (lowercase) boldface letters denote matrices (col-
umn vectors). =√−1 denotes the imaginary unit. ‖X‖1
and ‖x‖1 denote the ℓ1-norm of X and x, and ‖x‖2 denotes
the ℓ2-norm of x. ⌈x⌉ denotes the smallest integer that is
not less than x, while ⌊x⌋ denotes the largest integer that
is not greater than x. ⊙ is the Hadamard product operator,
and ⊗ is the Kronecker product operator. (·)M denotes mod
M , and 〈x〉N denotes(x+
⌊N2
⌋)
N−⌊N2
⌋. The notation
∆= is
used for definitions, and ≅ is used to indicate equivalence
or approximate equivalence. δ (·) denotes the Dirac delta
function. [X]a,: denotes the a-th row of X, wihle [X]:,bdenotes the b-th colomn of X. [X]a,b denotes the (a, b)-thelement of X. [x]a denotes the a-th element of x.
II. SYSTEM MODEL
In this section, some basic definitions and concepts of OTFS
are reviewed first. Then we analyze the input-output relation-
ship of the SISO OTFS system, where the OFDM modem is
selected as time-frequency modem, and the fractional Doppler
is considered. Then we extend it to the case of massive MIMO
systems.
3
A. SISO OTFS System
For the sake of the subsequent derivation, we first define a
lattice in the time-frequency domain as Λ = {nTsym,m∆f},
where Tsym (seconds) and ∆f (Hz) are sampling intervals of
the time-dimension and the frequency-dimension, respectively,
n = 0, · · · , N − 1, m = 0, · · · ,M − 1, N,M ∈ Z+.
Similarly, the lattice in the delay-Doppler domain is defined
as Γ ={
kNTsym
, lM∆f
}
where 1M∆f
and 1NTsym
are sampling
intervals of the delay-dimension and Doppler-dimension, re-
spectively, and k =⌈−N
2
⌉, · · · ,
⌈N2
⌉− 1, l = 0, · · · ,M − 1.
We then arrange a set of NM quadrature amplitude mod-
ulated (QAM) symbols XDD [k, l] in the delay-Doppler do-
main on the lattice Γ. The inverse symplectic finite Fourier
transform (ISFFT) is utilized at OTFS transmitter to convert
XDD [k, l] to the symbols XTF [n,m] in the time-frequency
domain as [32]
XTF[n,m]=1√MN
⌈N/2⌉−1∑
k=⌈−N/2⌉
M−1∑
l=0
XDD [k, l] e−2π(mlM
−nkN ).
(1)
Next, an OFDM modulator can be used as a time-frequency
modulator to convert XTF [n,m] to a transmitted signal s(t)with a transmitted waveform gtx(t) as
s(t) =
M−1∑
m=0
N−1∑
n=0
XTF[n,m]e2πm∆f(t−MCPT
M−nTsym)
× gtx(t− nTsym), (2)
where MCP is the length of CP, T =MTsym
M+MCPis time duration
of an OFDM symbol without CP, and gtx(t) is defined as
gtx(t)∆=
{ 1√T, 0 ≤ t ≤ Tsym
0, otherwise. (3)
Note that the CP is added in this step.
After the transmitted signal s(t) passing through the mul-
tipath time-variant channel, the received signal r(t) can be
obtained as
r(t) =
∫∫
hc(τ, ν)e2πν(t−τ)s(t− τ)dτdν + n(t), (4)
where τ is the delay, ν is the Doppler frequency, hc(τ, ν) is
impulse response in the delay-Doppler domain [33], and n(t)is the additive Gaussian noise. Usually, there are only a few
scatterers in the transmission environment. Thus the channel
can be represented in a sparse way [14]. We assume that the
number of the dominant paths between the transmitter and the
receiver is P , and each consists of S subpaths. Hence, hc(τ, ν)is given by
hc(τ, ν) =
P−1∑
i=0
S−1∑
si=0
hsiδ (τ − τi)δ (ν − νsi) , (5)
where hsi and νsi are the complex path gain and the Doppler
frequency of the (si+1)-th subpath of the (i+1)-th dominant
path, respectively. All of subpaths in the (i + 1)-th dominant
path have the same delay τi [34]. We define the delay and
Doppler taps for si subpath as
τi =li
M∆f, νsi =
ksi + ksiNTsym
, (6)
where li and ksi are integers and represent the indexes of delay
and Doppler taps. The real number ksi , whose value range is(− 1
2 ,12
], is defined as the fractional Doppler. The fractional
delay is not considered since the resolution 1M∆f
of the delay
axis is sufficient so that each path delay can be mapped to an
integer delay tap in typical wide-band systems [35].
At the receiver, the cross-ambiguity function between a
received waveform grx(t) and the received signal r(t) is
computed by the matched filter as
Agrx,r (τ, ν) =
∫
e−2πν(t−τ)g∗rx(t− τ)r(t)dt, (7)
where grx(t) is defined as
grx(t)∆=
{ 1√T, MCPT
M≤ t ≤ Tsym
0, otherwise. (8)
Note that the CP is removed in this step. By sampling
the cross-ambiguity function, the received data in the time-
frequency domain is given by
Y TF [n,m] = Agrx,r (τ, ν)|τ=nTsym,ν=m∆f. (9)
It is worth noting that when the time duration of CP, i.e., MCPTM
is beyond the maximum path delay of all dominant paths, there
is no ISI between OFDM symbols within an OTFS symbol at
the receiver, which is different from [13]–[15].
Next, the symplectic finite Fourier transform (SFFT) can be
utilized to map Y TF [n,m] to the symbols Y DD [k, l] in the
delay-Doppler domain as
Y DD [k, l] =1√NM
N−1∑
n=0
M−1∑
m=0
Y TF [n,m] e2π(mlM
−nkN ).
(10)
We define the complex gain of the time-variant channel on the
delay tap l at time ρTs as
hρ,l =
P−1∑
i=0
S−1∑
si=0
hsie2π(ρ−l)Tsνsi δ (lTs − τi), (11)
where Ts = 1M∆f
is the system sampling interval, and we
define hsiρ,li
as the complex gain of the (si + 1)-th subpath
of hρ,li . Through the above discussion, we give the input-
output relationship of the SISO OTFS system, as shown in
Proposition 1.
Proposition 1: For a SISO OTFS system, when the OFDM
modem is used as the time-frequency modem, the input-output
relationship is expressed as
Y DD [k, l] =1√N
P−1∑
i=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉HDD [k′, li, l]
×XDD [〈k − k′〉N , (l− li)M ] + ZDD [k, l] ,(12)
where
HDD [k′, li, l]
∆=
1√N
N−1∑
j=0
(S−1∑
si=0
hsiMCP+j(M+MCP),li
e2π
l(ksi+ksi )(M+MCP)N
)
e−2π k′jN ,
(13)
4
and ZDD [k, l] is the additive noise in the delay-Doppler
domain.
Proof : See Appendix.
From (13), we have the following two findings. First,
HDD [k′, li, l] is related to l, which is the received data
position along the delay-dimension. Second, for the (i+1)-thdominant path, HDD [k′, li, l] is affected by all subpaths due
to the fractional Doppler. These two points make the system
model very complicated, which is not conducive to subsequent
analysis. Hence, we utilize the characteristics of subpaths to
simplify it. Since all subpaths of one dominant path usually
originate from the same scattering cluster, the directions of
arrival deviation of these subpaths at the user terminal are
usually slight [36]. Therefore, we approximate the Doppler
frequency of all subpaths of the (i + 1)-th dominant path
to the same value νi = ki+ki
NTsym, i.e., ksi + ksi ≈ ki + ki,
si = 0, · · · , S−1. Such an approximation simplifies the model
in (12) to the follows
Y DD [k, l] ≅1√N
P−1∑
i=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉HDD [k′, li]e
2πl(ki+ki)
(M+MCP)N
×XDD [〈k − k′〉N , (l − li)M ] + ZDD [k, l] ,(14)
where HDD [k, l] is the delay-Doppler domain channel, which
is defined as
HDD [k, l] =1√N
N−1∑
j=0
hMCP+j(M+MCP),le−2π kj
N . (15)
Equality in (14) holds precisely if there is only one subpath
in a dominant path (i.e., S = 1).
It is noticed that when N → ∞, (12) is transformed into
Y DD[k, l]N→∞=
1√N
M−1∑
l′=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉HDD [k′, l′]e
2π lk′
(M+MCP)N
×XDD[〈k − k′〉N , (l − l′)M ]+ZDD [k, l] ,(16)
which is the same as that in [28] and [29]. Therefore, the model
in [28] and [29] can be seen as a special case of our proposed
input-output relationship when N → ∞. Moreover, comparing
(14) with (16), we can find that the phase compensation
e2π
l(ki+ki)(M+MCP)N in (14) uses the information of each dominant
path, which makes the model more accurate in the practical
system.
Note that we choose the OFDM-based OTFS system similar
to [11] in this paper, which can be achieved simply as an
overlay of the widely used OFDM system, instead of the
spectral efficient OTFS model in [13]–[15]. If fewer CPs are
used to improve spectral efficiency as in [13]–[15], the input-
output relationship will have an additional case owing to the
ISI between OFDM symbols within an OTFS symbol, and the
only difference between the two cases is the exponential term,
which implies that the basic idea of using the channel path
information to modify the phase compensation matrix of the
sensing matrix to improve the channel estimation performance,
as shown later, is still valid in the spectral efficient OTFS
model.
B. Massive MIMO-OTFS System
We consider the downlink transmission in a massive MIMO-
OTFS system. The BS deploys Nt antennas and serves Usingle-antenna user terminals. We consider the centralized
massive MIMO, where the uniform linear array (ULA) is
equipped at the BS, and the antenna spacing is set to half
wavelength. Without loss of generality, we focus on one user
terminal and disregard the dependency of the channel on user
index for simplicity.
Similar to (14), the symbols received at the user terminal
in the delay-Doppler domain can be expressed as
Y DD [k, l]
≅1√N
Nt−1∑
nt=0
P−1∑
i=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉HDDS [k′, li, nt] e
2πl(ki+ki)
(M+MCP)N
×XDDS [〈k − k′〉N , (l − li)M , nt] + ZDD [k, l] , (17)
where XDDS [k, l, nt] is transmitted symbols in the delay-
Doppler-space domain, HDDS [k′, li, nt] is the delay-Doppler-
space domain channel, which is defined as [37]
HDDS [k, l, nt]
=1√N
N−1∑
j=0
P−1∑
i=0
S−1∑
si=0
hsie2π(MCP+j(M+MCP)−l)Tsνsi
× δ (ℓTs − τi) eπnt sinϕsi e−2π kj
N , (18)
where ϕsi ∈ [−π/2, π/2) is the angle of departure (AoD) of
the (si+1)-th subpath. To exploit the sparsity of the beam do-
main, the delay-Doppler-beam domain channel HDDB [k, l, b]is obtained by applying normalized discrete Fourier transform
(DFT) for the delay-Doppler-space domain channel along the
space-dimension nt as [28], [38]
HDDB [k, l, b]
=1√Nt
Nt−1∑
nt=0
HDDS [k, l, nt]e−2π
bntNt
=
√N
Nt
P−1∑
i=0
S−1∑
si=0
hsie2π(MCP−l)TsνsiΞN (k −NTsymνsi)
× δ (lTs − τi) ΞNt (b−Nt sinϕsi/2) , (19)
where b = −Nt
2 , · · · , 0, · · · , Nt
2 − 1 is the beam index,
and ΞN (x)∆= 1
N
N−1∑
i=0
e−2π xN
i. From (19), we can find
that the dominant elements of HDDB [k, l, b] are distributed
in the positions where k ≈ NTsymνsi , l ≈ τiM∆f and
b ≈ Nt sinϕsi/2, which means that the channel has the 3D
sparsity over the delay-Doppler-beam domain [28], [39]. By
5
combining (17) and (19), the received symbols can be rewritten
as
Y DD [k, l]
≅1√N
Nt/2−1∑
b=−Nt/2
P−1∑
i=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉HDDB [k′, li, b] e
2πl(ki+ki)
(M+MCP)N
×XDDB [〈k − k′〉N , (l − li)M , b] + ZDD [k, l] , (20)
where
XDDB [k, l, b]∆=
1√Nt
Nt∑
nt=0
e2πrntNt XDDS [k, l, nt] . (21)
Equality in (20) holds precisely if there is only one subpath
in a dominant path. Similarly, when N → ∞, the received
symbols are given by
Y DD [k, l]
N→∞=
1√N
Nt/2−1∑
b=−Nt/2
M−1∑
l′=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉HDDB[k′, l′, b]e
2π lk′
(M+MCP)N
×XDDB [〈k − k′〉N , (l− l′)M , b] + ZDD [k, l] .(22)
According to downlink massive MIMO-OTFS system mod-
els and the sparsity of the delay-Doppler-beam domain chan-
nel HDDB [k, l, b], the downlink CSI acquisition is actually
a sparse signal reconstruction problem [28]. A variety of
compressive sensing (CS) algorithms can be used to solve
this problem. Next, we will give the model of downlink CSI
acquisition and propose the corresponding deterministic pilot
design and MSMCE algorithm.
III. DOWNLINK CSI ACQUISITION FOR MASSIVE
MIMO-OTFS SYSTEMS
In this section, we first model the downlink CSI acquisition
as a sparse signal reconstruction problem. To improve sensing
performance, the deterministic pilots are designed, and the
pilot overhead is also discussed. Then, we propose a MSMCE
algorithm to reconstruct the sparse delay-Doppler-beam do-
main channel and analyze its performance.
A. Downlink CSI Acquisition Model
Fig. 1 shows the position of pilots in the delay-Doppler
domain at one antenna. We consider that the position of pilots
in the delay-Doppler domain at each transmit antenna is the
same and given by
k = kp, · · · , kp +Np − 1, l = lp, · · · , lp +Mp − 1, (23)
where kp and lp are the initial position of the pilots in the
Doppler domain and delay domain, respectively, and Np and
Mp are the lengths of pilots. The guard intervals (i.e., zero
symbols) are demanded to eliminate the interference between
the data and pilots.
Delay-Doppler-beam domain channels have finite support
[0, τmax] and [−νmax, νmax] along the delay-dimension and
Doppler
Delay
Pilot Data Guard
0
M
pN
pM
gM
gM
2N-é ùê ú
g 2Né ùê ú g 2Né ùê ú
2 1N -é ùê ú
( ),p pl k
Fig. 1. The position of pilots, data, and guard intervals in the delay-Dopplerdomain at one antenna.
the Doppler-dimension, respectively [8]. Hence for the delay-
dimension, the guard intervals should be placed at the be-
ginning and the end of the pilots with the length Mg ≥τmaxM∆f . However, for the Doppler-dimension, νmaxNTsym
is typically non-integer in practical systems, which means that
according to the characteristic of ΞN (x) in (19), the magnitude
at each k is not zero and decreases as k moves away from
νmaxNTsym. Thus we only consider k ∈[−Nmax
2 , Nmax
2 − 1]
for HDDB and replace the values outside the range with zero,
and the guard intervals should be placed at the beginning
and the end of the pilots with the lengthNg
2 ≥ Nmax
2 .
Therefore, the range of l′ and k′ of the delay-Doppler-beam
domain channel HDDB [k′, l′, b] are limited to [0,Mg] and
[⌈−Ng/2⌉ , ⌈Ng/2⌉ − 1], respectively. Moreover, the data is
placed in the position except for the pilot and the guard
interval.
The position of received symbols for channel estimation is
the same as (23). Therefore, (20) can be rewritten as
yDD≅(Φ⊙XDDB
)hDDB + zDD, (24)
where XDDB ∈ CMpNp×MgNgNt is the pilot matrix
with element XDDB [〈k − k′〉N , (l − l′)M , b] of index
((l − lp)Np + (k − kp) + 1, (b+Nt/2)MgNg + l′Ng +k′ + ⌊Ng/2⌋ + 1), where k = kp, · · · , kp + Np − 1,
l = lp, · · · , lp + Mp − 1, b = −Nt
2 , · · · , 0, · · · , Nt
2 − 1,
k′ =⌈
−Ng
2
⌉
, · · · , 0, · · · ,⌈Ng
2
⌉
− 1 and l′ = 0, · · · ,Mg − 1.
In (24), yDD ∈ CMpNp×1 and zDD ∈ C
MpNp×1 are the vector
forms of received symbols and the additive noise with element
Y DD [k, l] and ZDD [k, l] of index ((l − lp)Np+(k − kp)+1).hDDB ∈ CMgNgNt×1 is the vector form of the delay-Doppler-
beam domain channel with element HDDB [k′, l′, b] of
index ((b + Nt/2)MgNg + l′Ng + k′ + ⌊Ng/2⌋ + 1),
6
Φ ∈ CMpNp×MgNgNt is the phase compensation matrix
with element φ (l, l′) of index ((l − lp)Np + (k − kp)+ ⌊Np/2⌋+ 1, (b+Nt/2)MgNg + l′Ng + k′+⌊Ng/2⌋+1),which is defined as
φ (l, l′) =
{
e2π
l(ki+ki)(M+MCP)N , l′ ∈ Del, l′ = li
1, l′ /∈ Del, (25)
where Del is the tap index set of the delay of all dominant
paths, i.e., Del = {l0, l1, · · · , lP−1}. ki + ki is selected from
the set Dop which consists of the integer and fractional
tap indexes of the Doppler frequency of all dominant paths,
i.e., Dop ={
k0 + k0, k1 + k1, · · · , kP−1 + kP−1
}
. When
N → ∞, according to (22), Φ in (24) is converted to
Φ ∈ CMpNp×MgNgNt with element φ (l, k′) = e2π lk′
(M+MCP)N
of index ((l − lp)Np + (k − kp) +1, (b + Nt/2)MgNg +l′Ng + k′+⌊Ng/2⌋+1). And in this case, the model is the
same as that in [28].
We can rewrite (24) as
yDD≅ ΘhDDB + zDD, (26)
where Θ∆= Φ⊙XDDB. Thus the downlink CSI acquisition of
massive MIMO-OTFS systems is converted to a sparse signal
reconstruction problem, where hDDB is the sparse vector to
be recovered, and Θ is the sensing matrix. When N → ∞, Θ
is converted to Θ∆= Φ⊙XDDB. Hence we have two kinds of
sensing matrix that can be used in CS, and the only difference
between them is the phase compensation matrix. The phase
compensation matrix Φ is more accurate but requires both path
delay and Doppler frequency of each dominant path, which can
not be directly obtained. In contrast, the phase compensation
matrix Φ is irrelevant to the channel, although it is inaccurate.
To combine the advantages of both, the MSMCE algorithm is
proposed and will be discussed in detail later. Prior to that,
we give the deterministic pilot design in the next subsection.
B. Deterministic Pilot Design
In order to recover the sparse vector reliably, the sensing
matrix must be designed carefully and satisfies the restricted
isometry property (RIP) proposed in [40]. The RIP implies
that the coherence (i.e., inner product) between columns in
a sensing matrix should be as small as possible to obtain a
good sensing performance. Although the random matrix is
usually used as the sensing matrix due to its near-optimality
[41], it is hard to be implemented and costs a lot of memory
consumption for its randomness. Therefore, we focus on
the deterministic pilot design for practical communication
systems.
Since the phase compensation matrix Φ of the sensing
matrix Θ is related to the channel, which varies with time,
the pilot matrix XDDB is what we analyzed exactly. The
design of pilot matrix XDDB is equivalent to the design of
transmitted pilots XDDB [k, l, b] in the delay-Doppler-beam
domain with their position in the delay-Doppler domain at
each beam. Note that the pilots XDDS [k, l, nt] in the delay-
Doppler-space domain (i.e., the pilots in the delay-Doppler
domain at each BS transmit antenna) can be obtained using
(21), and such a transformation will not affect the position of
pilots in the delay-Doppler domain.
We first analyze the relationship between the pilot matrix
XDDB and the pilot position at one beam. As shown in Fig.
2(a), each dashed box contains MpNp pilot symbols, which
are used to construct the corresponding column of the pilot
matrix in Fig. 2(b). The pilots at each beam can construct
MgNg columns totally.
pNg 2Né ùê ú g 2Né ùê ú
pM
gM
gM
( ),p pl k
(a)
tg g 1
2
Nb M Næ ö+ +ç ÷
è ø
tg g g
2
Nb M N Næ ö+ +ç ÷
è ø
tg g g g
2
Nb M N M Næ ö+ +ç ÷
è ø
(b)
Fig. 2. The relationship between the position of pilots at one beam andthe pilot matrix. (a) Position of pilots in the delay-Doppler domain at the(
b+ Nt2
+ 1)
-th beam; (b) Pilot matrix.
However, we can find that if we directly use the position of
pilots shown in Fig. 1, there will be some zero symbols in the
pilot matrix, which will affect the coherence between columns
of the pilot matrix and make the pilot design difficult. There-
fore, based on Fig. 1, we propose a new pilot position design
shown in Fig. 3. Specifically, along the Doppler-dimension,
the first ⌈Ng/2⌉ pilots are added to the end (marked in yellow
grids in Fig. 3), and the last ⌈Ng/2⌉ pilots are added to the
beginning (marked in pink grids in Fig. 3) while the guard
intervals are placed at the beginning and the end of pilots
7
Doppler
Delay
Pilot Data Guard
0
M
pN
pM
gM
gM
gM
2N-é ùê ú
g 2Né ùê ú g 2Né ùê ú g 2Né ùê ú g 2Né ùê ú
2 1N -é ùê ú
( ),p pl k
Fig. 3. The designed position of pilots, data, and guard intervals in the delay-Doppler domain at one beam.
with the length ⌈Ng/2⌉. Along the delay-dimension, the last
Mg pilots are added to the beginning (marked in thick-black
grids in Fig. 3) while the guard intervals are placed at the end
of the pilots with the length of Mg. The data is placed in the
position except for the pilot and the guard interval.
Next, we discuss the pilot design at a given beam and
between different beams. For a given beam, we assume that
the ((b+ Nt
2
)MgNg + 1)-th column of the pilot matrix (i.e.,
the black dashed box in Fig. 2(b)) constructed by the pilots at
this beam shown in Fig. 3 is defined as
black : c(b+Nt2 )MgNg+1
= pMp
0 ⊗ pNp
0 , (27)
where pLc denotes a sequence of length L and is cyclically
shifted by c symbols. According to the designed pilot position
and its relationship with the pilot matrix, different columns
(i.e., dashed boxes with different colors in Fig. 2(b)) can be
expressed as
red : c(b+Nt2 )MgNg+2
= pMp
0 ⊗ pNp
1
green : c(b+Nt2 )MgNg+Ng
= pMp
0 ⊗ pNp
Ng−1
purple : c(b+Nt2 )MgNg+Ng+1
= pMp
1 ⊗ pNp
0
brown : c(b+Nt2 )MgNg+MgNg
= pMp
Mg−1 ⊗ pNp
Ng−1
. (28)
Therefore, according to the property of Kronecker product
(A⊗B) (C⊗D) = (AC)⊗(BD), the pilot sequence should
be orthogonal to its cyclically shifted version to ensure the
orthogonality between columns of the pilot matrix at the given
beam, which is defined as the orthogonality condition.
The above pilot design can be used at up to ηdelηdop beams,
where ηdel = ⌊Mp/Mg⌋ and ηdop = ⌊Np/Ng⌋. The difference
between these beams is that pMp
0 should be replaced with
pMp
iMg, or p
Np
0 with pNp
jNg, where 1 ≤ i ≤ ηdel−1 and 1 ≤ j ≤
ηdop−1. Note that at least one of the two pilot sequences needs
to be replaced to ensure the orthogonality between columns.
However, for the beam other than these ηdelηdop beams, a new
pair of pilot sequences should be used. And the coherence of
the columns constructed by this new pair of pilot sequences
and the columns constructed by other pairs of pilot sequences
should be as low as possible to ensure the sensing performance
of the entire pilot matrix, which is defined as the low coherence
condition.
To sum up, the designed deterministic pilots should satisfy
both the orthogonality condition and the low coherence condi-
tion. ZC sequence is quite suitable for the proposed determin-
istic pilot design since it is a cyclic orthogonal sequence with
good autocorrelation and low cross-correlation. We define the
ZC sequence cyclically shifted by c symbols with root r and
length L as zL,rc , and its (k+1)-th element can be expressed
as
zL,rc [k] =
1√Le
πr(k−c)L((k−c)L+(L)2)L , (29)
where k = 0, 1, · · · , L − 1. Note that L in the current
system should be Mp or Np. The cyclic orthogonality of
the ZC sequence makes it satisfy the orthogonality condition.
Meanwhile, when the length of ZC sequence L is prime,
the magnitude of cross-correlation between two normalized
ZC sequences with different roots is 1√L
[42]. It means that
when we need a new pair of pilot sequences, a pair of
ZC sequences with a new pair of roots can be chosen, and
they satisfy the low coherence condition. Therefore, based on
the ZC sequence, the proposed deterministic pilot design is
summarized in Algorithm 1, where the pilots at all beams are
rearranged as a matrix XP of size MpNp ×Nt.
Algorithm 1: Deterministic pilot design
Input : Mp, Np, Mg, Ng
Output: XP
1 Initialize: ηdel = ⌊Mp/Mg⌋, ηdop = ⌊Np/Ng⌋, g = 0,
2 b = 0, γ = Mp − 1, µ = Np − 13 while gηdelηdop < Nt do
4 for i = 0 to ηdel − 1 do
5 for j = 0 to ηdop − 1 do
6 b = gηdelηdop + iηdop + j + 17 if b > Nt then
8 break
9 end
10
[
XP]
:,b= z
Mp,γ
iMg⊗ z
Np,µ
jNg
11 end
12 end
13 g = g + 1, γ = γ − 1, µ = µ− 114 end
15 return XP
To analyze the correlation between the columns of the
designed deterministic pilot matrix, we divide them into g sets
with MgNgηdelηdop columns in each set. The last set contains
8
MgNgηdelηdop columns where ηdel ≤ ηdel and ηdop ≤ ηdop.
Hence the deterministic pilot matrix XDDB is given by
group1︷ ︸︸ ︷
c1,1, · · · c1,MgNgηdelηdop, · · · ,
groupg︷ ︸︸ ︷
cg,1, · · · cg,MgNgηdelηdop
.
(30)
Thus the coherence between the columns of the pilot matrix
XDDB can be given by
∣∣cHi1,j1 ci2,j2
∣∣ =
1, i1 = i2, j1 = j20, i1 = i2, j1 6= j2
1√MpNp
, i1 6= i2. (31)
Note that the maximal coherence between the columns
of the pilot matrix has a lower bound√
MgNgNt−MpNp
(MgNgNt−1)MpNp,
which is called Welch bound [43]. In massive MIMO-OTFS
systems, MgNgNt is usually much larger than MpNp, which
means that the maximal coherence between the columns of
the designed deterministic pilot matrix shown in (31) is very
close to Welch bound. Moreover, the designed deterministic
pilot can be quickly generated according to different system
configurations instead of spending a lot of memory for storage
in advance, while the random pilot is hard to be implemented
and consumes much memory due to its randomness.
Since the sensing matrix is the Hadamard product of pilot
matrix XDDB and phase compensation matrix Φ or Φ, the
correlation between the columns of the sensing matrix is
affected by the phase compensation matrix. Nevertheless, the
proposed deterministic pilot design still has the advantage over
the random pilot, as shown in simulation results.
C. Pilot Overhead
Comparing Fig. 3 with Fig. 1, we can find that the delay-
Doppler domain grids occupied by pilots and guard intervals
of the proposed pilot design is Ng (Mp + 2Mg) more than
that of conventional random pilots used in [28] when the
same number of pilots (i.e., MpNp) is considered. However,
the proposed pilot design can save about 5-10 percent of the
number of pilots compared with random pilots due to the
better sensing performance, as shown in simulation results,
which means that the overall pilot overhead of the proposed
pilot design is less. Moreover, when N is not large in practical
systems, the fractional Doppler will cause severe inter-Doppler
interference, which means that pilots usually need to occupy
the entire Doppler domain (i.e., Np = N ) to prevent interfer-
ence between the data and pilots. In this case, guard intervals
along the Doppler-dimension are no longer required, and only
the last Mg pilots are added to the beginning along the delay-
dimension to construct the structure of Fig. 3. Therefore, the
delay-Doppler domain grids occupied by pilots and guard
intervals shown in Fig. 3 are the same as that of Fig. 1,
which means that the overall saved pilot overhead can further
increase.
D. MSMCE Algorithm
In this subsection, we will give a detailed discussion of the
proposed MSMCE algorithm, which utilizes both two sensing
matrices Θ and Θ to solve the sparse signal recovery problem.
The delay-Doppler-beam domain channel hDDB can be
recovered by the CS algorithm using the sensing matrix
and received symbols. Usually, N can not be too large in
the practical system, which implies that the performance of
the downlink CSI acquisition based on sensing matrix Θ
is unsatisfactory due to the inaccurate sensing matrix when
the signal-to-noise ratio (SNR) is high. Therefore, we prefer
to use the sensing matrix Θ, which is more accurate for
channel estimation. However, according to (25), the phase
compensation matrix Φ in the sensing matrix Θ depends on
the information of each path. Therefore, when there is no
knowledge of path delays and Doppler frequencies of the
channel at first, we can choose Θ as the sensing matrix to
estimate the channel with CS algorithms since Θ is irrelevant
to the channel. Then we extract the path delay and Doppler
frequency of each dominant path from the channel estimation
results hDDB to build a more accurate sensing matrix Θ.
Hence, before introducing the MSMCE algorithm, we propose
a method to extract the path delays and Doppler frequencies
from the channel estimation results, which is crucial in the
MSMCE algorithm.
Algorithm 2: The path delay and Doppler frequency
extraction
Input : hDDB, εOutput: Del, Dop
1 Initialize: Del = ∅, Dop = ∅2 Rearrange the magnitude of the channel estimation
output hDDB as HDDBmag ∈ CNg×Mg×Nt .
3 Sum the beam-dimension of the matrix HDDBmag to get
HDDmag ∈ CNg×Mg .
4 Sum the Doppler-dimension of the matrix HDDmag to
get hDmag ∈ CMg×1.
5 Σ =∥∥hD
mag
∥∥1
6 while∥∥hD
mag
∥∥1
/
Σ > ε do
7 lτ = argmaxl∈{1,2,··· ,Mg}
[hDmag
]
l
8 Del = {Del, lτ}9 kν1 = argmax
k∈{1,2,··· ,Ng}
[HDD
mag
]
k,lτ
10 kν2 = argmaxk∈{1,2,··· ,Ng}\{kν1}
[HDD
mag
]
k,lτ
11 kν =kν1−⌊Ng
2
⌋
−1, kν =(kν2−kν1)[H
DDmag]kν2 ,lτ
[HDDmag]kν1 ,lτ
+[HDDmag]kν2 ,lτ
12 Dop ={
Dop, kν + kν
}
13 Set the lτ -th element of hDmag to 0.
14 end
15 return Del, Dop
We first define the magnitude of the delay-Doppler-beam
domain channel as HDDBmag [k, l, b]
∆=∣∣HDDB [k, l, b]
∣∣. The
delay of each path can be easily obtained by summing the
Doppler-dimension and the beam-dimension of HDDBmag [k, l, b]
and searching positions with dominant values along the delay-
dimension. Hence, we focus on the extraction of the Doppler
frequency νi =ki+ki
NTsymof each dominant path. Note that we
9
approximate the Doppler frequency of all subpaths of each
dominant path to the same. Therefore, for the (i + 1)-thdominant path, HDDB
mag [k, li, b] can be expressed as
HDDBmag [k, li, b]
≅
√
N
Nt
∣∣∣∣∣
S−1∑
si=0
hsie2π(MCP−li)TsνsiΞNt(b−Nt sinϕsi/2)
∣∣∣∣∣
× |ΞN (k −NTsymνi)| . (32)
Then we sum the beam-dimension of HDDBmag [k, li, b] as
HDDmag [k, li] =
Nt/2−1∑
b=−Nt/2
HDDBmag [k, li, b]
≅ βi |ΞN (k −NTsymνi)| , (33)
where
βi =
√N
Nt
×Nt/2−1∑
b=−Nt/2
∣∣∣∣∣
S−1∑
si=0
hsie2π(MCP−li)TsνsiΞNt(b−Nt sinϕsi/2)
∣∣∣∣∣.
(34)
Note that |ΞN (k −NTsymνi)| in (33) reaches the maximum
at k = ki + ki and decreases as k moves away from ki + ki.Hence, we define that
kν1 = argmaxk∈{⌈−Ng/2⌉,··· ,⌈Ng/2⌉−1}
HDDmag [k, li] ,
kν2 = argmaxk∈{⌈−Ng/2⌉,··· ,⌈Ng/2⌉−1}\{kν1}
HDDmag [k, li] . (35)
It can be checked that the integer part of the Doppler frequency
is ki = kν1 , |kν2 − kν1 | = 1, and ki+ ki must be between kν1and kν2 . Then we can get that
HDDmag [kν1 , li]
HDDmag [kν2 , li]
=
∣∣∣ΞN (−ki)
∣∣∣
∣∣∣ΞN (kν2 − kν1 − ki)
∣∣∣
(a)=
∣∣∣∣∣∣
sin(
−kiπ)
sin(
−ki
Nπ)
∣∣∣∣∣∣
·
∣∣∣∣∣∣∣∣
sin((
kν2 − kν1 − ki
)
π)
sin
(
(kν2−kν1−ki)N
π
)
∣∣∣∣∣∣∣∣
−1
,
(36)
where (a) follows from the fact that |ΞN (x)| = 1N
∣∣∣∣sin(πx)
sin( πxN )
∣∣∣∣.
Since |kν2 − kν1 | = 1, which means that
∣∣∣sin
(
−kiπ)∣∣∣ =
∣∣∣sin
((
kν2 − kν1 − ki
)
π)∣∣∣, (36) can be rewritten as
HDDmag [kν1 , li]
HDDmag [kν2 , li]
=
∣∣∣∣sin
(
(kν2−kν1−ki)N
π
)∣∣∣∣
∣∣∣sin
(−ki
Nπ)∣∣∣
≈
∣∣∣
(
kν2 − kν1 − ki
)∣∣∣
∣∣∣−ki
∣∣∣
. (37)
Therefore, the fractional Doppler can be derived from (37) as
ki =(kν2 − kν1)H
DDmag [kν2 , li]
HDDmag [kν1 , li] +HDD
mag [kν2 , li]. (38)
Similarly, the tap index set of the path delay Del and Doppler
frequency Dop of all dominant paths are obtained, and the
process is summarized in Algorithm 2, where ε is used
to terminate the iteration and usually a smaller value. Note
that there is no interference between the dominant paths in
Algorithm 2, which is because that the fractional delay is not
considered in this paper.
Next, the proposed MSMCE algorithm is given in Algo-
rithm 3. If we have no information about the channel to be
estimated at the beginning, Θ should be used as the sensing
matrix to obtain a temporary channel estimation result by
the CS algorithm. Then we use the initial estimation result
to extract the path delays and Doppler frequencies of all
dominant paths, which are used to modify the inaccurate
sensing matrix to obtain a more accurate sensing matrix Θ.
Finally, the channel estimation result hDDB is obtained by the
CS algorithm based on Θ, and Del and Dop are also updated
for the next channel estimation. Since the channel geometry
changes very slowly relative to the communication timescale
[9], in the next channel estimation, Del and Dop can be directly
used to construct the modified sensing matrix to estimate the
channel.
Algorithm 3: MSMCE Algorithm
Input : XDDB, yDD, Del, Dop
Output: hDDB, Del, Dop
1 if Del = ∅ or Dop = ∅ then
2 Sensing martix Θ = Φ⊙XDDB
3 Calculate the channel estimation result hDDB by
the CS algorithm based on Θ and yDD.
4 Calculate the Del and Dop by the Algorithm 2
based on hDDB.5 end
6 Use Del and Dop to construct the phase compensation
matrix Φ.
7 Sensing martix Θ = Φ⊙XDDB
8 Calculate the channel estimation result hDDB by the
CS algorithm based on Θ and yDD.
9 Update the Del and Dop by the Algorithm 2 based on
hDDB.
10 return hDDB, Del, Dop
E. The Performance Analysis of the MSMCE Algorithm
The complexity of the proposed MSMCE algorithm mainly
comes from two aspects: the Algorithm 2 and the selected CS
algorithm. We can find that most of the operations in the iter-
ation process of Algorithm 2 are comparisons, the number of
which is about P (Mg + 2Ng − 1), where P is the number of
the dominant paths. Therefore, the complexity of Algorithm 2
is very low compared with the CS algorithm, which means that
10
the MSMCE algorithm can obtain a significant performance
gain at the cost of a negligible increase in the complexity.
However, note that the MSMCE algorithm is based on the
assumption that we approximate the Doppler frequency of
all subpaths of each dominant path to the same value νi =ki+ki
NTsym. We now discuss the influence of the approximation
error caused by this assumption on the proposed MSMCE
algorithm.
Recalling (12), the approximation error can be defined as
follows
ζ =
P−1∑
i=0
S−1∑
si=0
∣∣∣∣∣e2π
l(ksi+ksi−ki−ki)
(M+MCP)N − 1
∣∣∣∣∣. (39)
We can find that if each dominant path contains only one
subpath, there is no approximation error, i.e., ζ = 0, and
the MSMCE algorithm achieves the optimal performance.
However, if more subpaths are considered, ζ will increase and
lead to a decrease in the accuracy of the derived downlink
CSI acquisition model and the extracted path information, and
eventually make the performance of the MSMCE algorithm
degrades.
Such an approximation error is related to several factors.
First, the user velocity will affect ζ. This is because that
although the directions of arrival deviation of subpaths are
very slight, too high the velocity will still make the Doppler
frequency difference between subpaths become larger and
increase the approximation error. Second, it is noticed that l in
(39) is the position of pilots along the delay-dimension, which
implies that the pilot position will have an influence on the
channel estimation performance. Specifically, the larger the l,the more severe the approximation error. Therefore, the initial
pilot position along the delay-dimension lp should be as close
to zero as possible to obtain an accurate CSI. Note that the
pilot position along the Doppler-dimension will not affect the
performance since it is irrelevant to the approximation error
shown in (39).
Moreover, for the proposed MSMCE algorithm, there are
two special cases that we need to discuss. First, when Nis large enough that the resolution of the Doppler domain
is sufficient to map the Doppler frequency of each path to
an integer tap, the model described in (22) is accurate in
this case, and the phase compensation matrix Φ is exact.
Therefore, there is no need to extract the path delays and
Doppler frequencies from the previous channel estimation
results to modify the sensing matrix, and only steps 2-3 in
the MSMCE algorithm are needed to estimate the channel.
Second, when the channel geometry changes very rapidly or
the time intervals between the channel estimations are too
long, the path information Del and Dop extracted from the
previous channel estimation results is not accurate enough,
which leads to the degradation of the MSMCE algorithm
performance. In this case, we need to re-extract the path
information with the current channel estimation result and use
the extracted information to construct the modified sensing
matrix to obtain a more accurate CSI.
TABLE ISIMULATION PARAMETERS
Parameter Values
Carrier frequency (GHz) 4
Subcarrier spacing (kHz) 15
Size of OTFS symbol (M,N) (512, 19 / 11 / 7)
Length of CP MCP 128
Number of BS antennas 16 / 32 / 64
Number of user terminal antennas 1
Scenario Urban macro cell
The number of dominant paths 6
The number of subpaths per dominant path 1 / 5 / 10 / 20
User velocity (m/s) 10 ∼ 130
IV. SIMULATION RESULTS
In this section, we utilize the numerical simulation to
illustrate the performance of the proposed downlink CSI acqui-
sition scheme. We choose the 3D-SOMP algorithm proposed
in [28] as the CS algorithm used in MSMCE. The performance
of the initial sensing matrix based channel estimation (ISMCE)
is also presented for comparison, where only Θ is used as the
sensing matrix (i.e., the steps 2-3 in the MSMCE algorithm).
The normalized mean square error (NMSE) is computed as
NMSE =
∥∥∥hDDB − hDDB
∥∥∥
2
2
‖hDDB‖22. (40)
We use the quasi deterministic radio channel generator
(QuaDRiGa) to generate over 105 channel realizations for
the Monte-Carlo simulations [34]. The relevant simulation
parameters are given in Table I. Note that since N is not
large, the sparsity of the delay-Doppler-beam domain channel
along the Doppler-dimension is not obvious. Therefore, we
set NP = N to prevent the interference between the data and
pilots. Unless explicitly mentioned otherwise, the pilots used
are the proposed deterministic pilots. The pilot overhead ratio
is defined as the ratio between the number of pilots and total
delay-Doppler domain resource grids, i.e.,MpNp
MN.
In Fig. 4, we compare the NMSE performance of different
algorithms with different types of pilots against the pilot
overhead ratio. The random pilots consist of complex Gaussian
random sequences. The number of BS antennas is 16, the
SNR is 5dB, N = 19, the user velocity is 100 m/s, and
the number of subpaths per dominant path is 20. We observe
that the proposed MSMCE algorithm outperforms the ISMCE
algorithm under different pilot overhead ratios when the same
type of pilots is utilized. This is because that the MSMCE
algorithm considers the fractional Doppler and uses the in-
formation extracted from the channel to reconstruct a more
accurate sensing matrix. Moreover, the proposed deterministic
pilot design has a better performance than the conventional
random pilots when the same algorithm is used. For example,
when the MSMCE algorithm is used, only 35% pilot overhead
ratio is required for the proposed deterministic pilot design
to achieve the NMSE of 0.013, while random pilots need at
least 45% pilot overhead ratio. This is due to the fact that the
coherence between the columns of the designed pilot matrix
is very low and has a better sensing performance compared
with random pilots.
11
0.2 0.25 0.3 0.35 0.4 0.45 0.5
pilot overhead ratio
10-3
10-2
10-1N
MS
E
MSMCE, deterministicMSMCE, randomISMCE, deterministicISMCE, random
Fig. 4. The NMSE performance versus pilot overhead ratio. The number ofBS antennas is 16, the SNR is 5 dB, and the user velocity is 100 m/s.
-10 -5 0 5 10 15
SNR(dB)
10-3
10-2
10-1
100
NM
SE
ISMCE, S=20, deterministicISMCE, S=5, deterministicMSMCE, S=20, deterministicMSMCE, S=5, deterministicISMCE, S=20, randomISMCE, S=5, randomMSMCE, S=20, randomMSMCE, S=5, random
Fig. 5. The NMSE performance versus SNR and comparison of the MSMCEalgorithm and the ISMCE algorithm. The number of BS antennas is 64, andthe user velocity is 100 m/s.
Fig. 5 shows the NMSE performance comparison of the
MSMCE algorithm and the ISMCE algorithm against the SNR,
where the number of subpaths per dominant path (i.e., S) is 20
or 5, and the pilots are random or deterministic. The number of
BS antennas is 64, the user velocity is 100 m/s, N = 19, and
the pilot overhead ratio is 89%. Such a high pilot overhead is
because the beam domain of the generated channel is not pure
sparse when the DFT matrix is used to convert the channel
from the space domain to the beam domain [44], which means
that more pilots are needed for both the ISMCE algorithm
and the MSMCE algorithm. We observe that our proposed
MSMCE algorithm has a substantial performance gain over
the ISMCE algorithm, regardless of the number of subpaths.
This is due to the fact that the MSMCE algorithm uses the
extracted path information to modify the sensing matrix to
-10 -5 0 5 10 15
SNR(dB)
10-4
10-3
10-2
10-1
100
NM
SE
S=20S=10S=5S=1
Fig. 6. The NMSE performance versus SNR and comparison of the MSMCEalgorithm with different numbers of subpaths per dominant path. The numberof BS antennas is 64, the user velocity is 100 m/s.
obtain more accurate CSI. Moreover, the channel estimation
with the proposed deterministic pilots outperforms that with
the random pilots due to the better sensing performance.
Fig. 6 presents the NMSE performance comparison of the
MSMCE algorithm with different numbers of subpaths per
dominant path. The number of BS antennas is 64, the user
velocity is 100 m/s, N = 19, and the pilot overhead ratio is
89%. It can be checked that with the decrease in the number of
subpaths per dominant path, the performance of the MSMCE
algorithm improves, and the optimal performance is obtained
when the number of subpaths is 1. This is because that the
less the number of subpaths, the smaller the approximation
error shown in (39). Moreover, the sensing matrix Θ is
completely accurate when each dominant path consists of only
one subpath since there is no approximation error.
In Fig. 7, we show the NMSE performance of the MSMCE
algorithm with different initial positions of pilots along the
delay-dimension, i.e., lp. The number of BS antennas is
32, the user velocity is 100 m/s, N = 19, and the pilot
overhead ratio is 49%. We observe that the performance of
the MSMCE algorithm degrades with an increase of lp when
the number of subpaths per dominant path is 20. This is
because that the approximation error increases with the lp and
makes the performance of the channel estimation degrades. We
can also observe that the slope of the curves increases with
the increment of SNR, which means that the impact of the
approximation error is more obvious in the high SNR regime.
However, when each dominant path contains only one subpath,
the performance is irrelevant to the position of pilots since
there is no approximation error in this case.
Fig. 8 shows the NMSE performance comparison at dif-
ferent user velocities. The number of BS antennas is 16, the
SNR is 15 dB, N = 19, the pilot overhead ratio is 35%,
and the number of subpaths per dominant path is 20. We
observe that the NMSE performance of the MSMCE algorithm
12
31 51 71 91 111 131 151 171 191 211 23110-3
10-2
10-1N
MS
E
SNR = 10 dB, S=20SNR = 15 dB, S=20SNR = 20 dB, S=20SNR = 15 dB, S=1
Fig. 7. The NMSE performance versus lp and comparison of the MSMCEalgorithm under different SNRs. The number of BS antennas is 32, and theuser velocity is 100 m/s.
0 20 40 60 80 100 120 140
User velocity (m/s)
10-3
10-2
NM
SE
MSMCEISMCE
Fig. 8. The NMSE performance versus user velocity. The number of BSantennas is 16, and the SNR is 15 dB.
degrades as the user velocity increases. This is because that
the difference between the Doppler frequencies of subpaths
increases with the user velocity, which leads to an increase in
the approximation error. However, the MSMCE algorithm still
has a significant performance gain over the ISMCE algorithm
at different user velocities.
In Fig. 9, we show the NMSE performance comparison un-
der different OTFS symbol sizes. The number of BS antennas
is 16, the user velocity is 100 m/s, the pilot overhead ratio
is 35%, and the number of subpaths per dominant path is
20. We can observe that as N decreases, the performance of
the ISMCE algorithm degrades, while the performance of the
MSMCE algorithm remains consistent. This is because that the
decrease in N will further reduce the resolution of the Doppler
domain, which makes the influence of the fractional Doppler
-10 -5 0 5 10 15
SNR(dB)
10-3
10-2
10-1
100
NM
SE
MSMCE, N=19MSMCE, N=11MSMCE, N=7ISMCE, N=19ISMCE, N=11ISMCE, N=7
Fig. 9. The NMSE performance versus SNR and comparison of the MSMCEalgorithm and the ISMCE algorithm under different OTFS symbol sizes. Thenumber of BS antennas is 16, and the user velocity is 100 m/s.
0 2.5 5 7.5 10 12.5 15 17.5 20
SNR(dB)
10-6
10-5
10-4
10-3
10-2
10-1
100B
ER
Perfect CSIMSMCEISMCE
Fig. 10. The BER performance versus SNR. The user velocity is 100 m/s,and the pilot overhead ratio is 35%.
more serious. Therefore, the ISMCE algorithm performs worse
under lower N due to the omission of fractional Doppler.
Finally, Fig. 10 shows the bit-error-rate (BER) performance
comparison of different channel estimation algorithms. The
transmitted symbols in the delay-Doppler domain are mod-
ulated by 4QAM. The user velocity is 100 m/s, N = 11,
the pilot overhead ratio is 35%, and the number of subpaths
per dominant path is 1. The minimum mean square error
(MMSE) detector is used to recover the transmitted data with
the CSI obtained by the MSMCE algorithm and the ISMCE
algorithm. Moreover, the BER performance of the perfect CSI
is provided for comparison. We can observe that the proposed
MSMCE algorithm can achieve satisfying BER performance,
which outperforms the ISMCE algorithm and is very close to
the perfect CSI case due to the consideration of the fractional
13
Doppler and the utilization of the extracted path information.
V. CONCLUSION
In this paper, we proposed a downlink CSI acquisition
scheme for massive MIMO-OTFS systems in presence of
the fractional Doppler, including deterministic pilot design
and channel estimation algorithm. Considering the fractional
Doppler in practical systems, we first analyzed the input-output
relationship of SISO OTFS systems based on the OFDM
modem and extended it to massive MIMO-OTFS systems.
Next, based on the downlink CSI acquisition model, the
deterministic pilot was designed, and the MSMCE algorithm
was presented to reconstruct the delay-Doppler-beam domain
channel. Simulation results demonstrated that the proposed
scheme could acquire accurate downlink CSI.
APPENDIX
PROOF OF PROPOSITION 1
By substituting (9) into (10), Y DD [k, l] can be expressed
as
Y DD [k, l] =1
NM
⌈N/2⌉−1∑
k′=⌈−N/2⌉
M−1∑
l′=0
αk,l [k′, l′]
×XDD [k′, l′] + ZDD [k, l] , (41)
where αk,l [k′, l′] is given by
αk,l [k′, l′]
=N−1∑
n=0
M−1∑
m=0
M−1∑
m′=0
e2π(m−m′)MCP
M e−2πn(k−k′)
N e2πml−m′l′
M
×∫∫
hc(τ, ν)Agrx ,gtx(−τ, (m−m′)∆f − ν)e−2πντ
× e−2π(m′∆fτ−nTsymν)dτdν
=
N−1∑
n=0
M−1∑
m=0
M−1∑
m′=0
e2π(m−m′)MCP
M e−2πn(k−k′)
N e2πml−m′l′
M
×P−1∑
i=0
S−1∑
si=0
hsi
1
M
M+MCP−1−li∑
p=MCP−li
e−2π(m′∆fτi−nTsymνsi)
× e−2πνsiτie−2π((m−m′)∆f−νsi)(p
M∆f+τi)
= NP−1∑
i=0
S−1∑
si=0
hsiΞN
(
k − k′ − ksi − ksi
)
×M
M+MCP−1−li∑
p=MCP−li
e2π
p(ksi+ksi )(M+MCP)N
× δ ((p−MCP + li − l)M ) δ ((p−MCP − l′)M ) ,(42)
where ΞN (x)∆= 1
N
N−1∑
i=0
e−2π xN
i, which also shows up after
(19).
Next, by substituting (42) into (41), Y DD [k, l] can be
written as
Y DD [k, l]
=
P−1∑
i=0
S−1∑
si=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉XDD [〈k − k′〉N , (l − li)M ]
×hsiΞN
(
k′ − ksi − ksi
)
e2π
(l−li+MCP)(ksi+ksi )(M+MCP)N +ZDD [k, l]
=1
N
P−1∑
i=0
⌈N/2⌉−1∑
k′=⌈−N/2⌉
N−1∑
j=0
S−1∑
si=0
XDD [〈k − k′〉N , (l − li)M ]
×hsiMCP+j(M+MCP),li
e2π
l(ksi+ksi )(M+MCP)N e−2π k′j
N +ZDD [k, l] ,
(43)
which completes the proof.
ACKNOWLEDGMENT
The authors would like to thank the associate editor and
the anonymous reviewers for their helpful comments and
suggestions.
REFERENCES
[1] J. Wu and P. Fan, “A survey on high mobility wireless communications:Challenges, opportunities and solutions,” IEEE Access, vol. 4, pp. 450–476, Mar. 2016.
[2] P. Fan, J. Zhao, and C. I., “5G high mobility wireless communications:Challenges and solutions,” China Commun., vol. 13, no. Supplement2,pp. 1–13, Feb. 2016.
[3] L. Cimini, “Analysis and simulation of a digital mobile channel usingorthogonal frequency division multiplexing,” IEEE Trans. Commun.,vol. 33, no. 7, pp. 665–675, Jul. 1985.
[4] G. L. Stuber, J. R. Barry, S. W. McLaughlin, Y. Li, M. A. Ingram,and T. G. Pratt, “Broadband MIMO-OFDM wireless communications,”Proc. IEEE, vol. 92, no. 2, pp. 271–294, Feb. 2004.
[5] Y. G. Li, J. H. Winters, and N. R. Sollenberger, “MIMO-OFDMfor wireless communications: Signal detection with enhanced channelestimation,” IEEE Trans. Commun., vol. 50, no. 9, pp. 1471–1477, Sep.2002.
[6] L. Deneire, P. Vandenameele, L. van der Perre, B. Gyselinckx, andM. Engels, “A low-complexity ML channel estimator for OFDM,” IEEE
Trans. Commun., vol. 51, no. 2, pp. 135–140, Feb. 2003.
[7] T. Wang, J. G. Proakis, E. Masry, and J. R. Zeidler, “Performancedegradation of OFDM systems due to Doppler spreading,” IEEE Trans.Wireless Commun., vol. 5, no. 6, pp. 1422–1432, Jun. 2006.
[8] R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Goldsmith, A. F.Molisch, and R. Calderbank, “Orthogonal time frequency space modu-lation,” in Proc. IEEE Wireless Commun. Netw. Conf., San Francisco,CA, USA, Mar. 2017, pp. 1–6.
[9] A. Monk, R. Hadani, M. Tsatsanis, and S. Rakib, “OTFS-orthogonaltime frequency space,” arXiv preprint arXiv:1608.02993, 2016.
[10] R. Hadani, S. Rakib, A. F. Molisch, C. Ibars, A. Monk, M. Tsatsanis,J. Delfeld, A. Goldsmith, and R. Calderbank, “Orthogonal time fre-quency space (OTFS) modulation for millimeter-wave communicationssystems,” in Proc. IEEE MTT-S Int. Microw. Symp., Honololu, HI, USA,Jun. 2017, pp. 681–683.
[11] A. Farhang, A. RezazadehReyhani, L. E. Doyle, and B. Farhang-Boroujeny, “Low complexity modem structure for OFDM-based orthog-onal time frequency space modulation,” IEEE Wireless Commun. Lett.,vol. 7, no. 3, pp. 344–347, Jun. 2018.
[12] K. R. Murali and A. Chockalingam, “On OTFS modulation for high-Doppler fading channels,” in Proc. IEEE Inf. Theroy Appl. Workshop,San Diego, CA, USA, Feb. 2018, pp. 1–10.
[13] P. Raviteja, K. T. Phan, and Y. Hong, “Embedded pilot-aided channelestimation for OTFS in delay-Doppler channels,” IEEE Trans. Veh.
Technol., vol. 68, no. 5, pp. 4906–4917, May 2019.
14
[14] P. Raviteja, K. T. Phan, Y. Hong, and E. Viterbo, “Interference can-cellation and iterative detection for orthogonal time frequency spacemodulation,” IEEE Trans. Wireless Commun., vol. 17, no. 10, pp. 6501–6515, Oct. 2018.
[15] Y. Ge, Q. Deng, P. C. Ching, and Z. Ding, “Receiver Design for OTFSwith fractionally spaced sampling approach,” IEEE Trans. WirelessCommun., Feb. 2021, early access.
[16] Z. Wei, W. Yuan, S. Li, J. Yuan, and D. W. K. Ng, “Transmitterand receiver window designs for orthogonal time frequency spacemodulation,” IEEE Trans. Commun., vol. 69, no. 4, pp. 2207–2223,Apr. 2021.
[17] M. Kollengode Ramachandran and A. Chockalingam, “MIMO-OTFS inhigh-Doppler fading channels: Signal detection and channel estimation,”in Proc. IEEE Global Commun. Conf., Abu Dhabi, United Arab Emi-rates, Dec. 2018, pp. 206–212.
[18] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,no. 11, pp. 3590–3600, Nov. 2010.
[19] L. You, X. Gao, G. Y. Li, X. Xia, and N. Ma, “BDMA for millimeter-wave/terahertz massive MIMO transmission with per-beam synchroniza-tion,” IEEE J. Sel. Areas Commun., vol. 35, no. 7, pp. 1550–1563, Jul.2017.
[20] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities andchallenges with very large arrays,” IEEE Signal Process. Mag., vol. 30,no. 1, pp. 40–60, Jan. 2013.
[21] L. You, X. Gao, A. L. Swindlehurst, and W. Zhong, “Channel acquisitionfor massive MIMO-OFDM with adjustable phase shift pilots,” IEEE
Trans. Signal Process., vol. 64, no. 6, pp. 1461–1476, Mar. 2016.[22] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE J. Sel.Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
[23] L. You, X. Gao, X. Xia, N. Ma, and Y. Peng, “Pilot reuse for massiveMIMO transmission over spatially correlated Rayleigh fading channels,”IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3352–3366, Jun.2015.
[24] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effi-ciency of very large multiuser MIMO systems,” IEEE Trans. Commun.,vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
[25] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “MassiveMIMO for next generation wireless systems,” IEEE Commun. Mag.,vol. 52, no. 2, pp. 186–195, Feb. 2014.
[26] L. You, M. Xiao, X. Song, Y. Liu, W. Wang, X. Gao, and G. Fettweis,“Pilot reuse for vehicle-to-vehicle underlay massive MIMO transmis-sion,” IEEE Trans. Veh. Technol., vol. 69, no. 5, pp. 5693–5697, May2020.
[27] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang,“An overview of massive MIMO: Benefits and challenges,” IEEE J.
Sel. Topics Signal Process., vol. 8, no. 5, pp. 742–758, Oct. 2014.[28] W. Shen, L. Dai, J. An, P. Fan, and R. W. Heath, “Channel estimation for
orthogonal time frequency space (OTFS) massive MIMO,” IEEE Trans.
Signal Process., vol. 67, no. 16, pp. 4204–4217, Aug. 2019.[29] Y. Liu, S. Zhang, F. Gao, J. Ma, and X. Wang, “Uplink-aided high mo-
bility downlink channel estimation over massive MIMO-OTFS system,”IEEE J. Sel. Areas Commun., vol. 38, no. 9, pp. 1994–2009, Sep. 2020.
[30] Y. Shan and F. Wang, “Low-complexity and low-overhead receiver forOTFS via large-scale antenna array,” IEEE Trans. Veh. Technol., Apr.2021, early access.
[31] M. Li, S. Zhang, F. Gao, P. Fan, and O. A. Dobre, “A new path divisionmultiple access for the massive MIMO-OTFS networks,” IEEE J. Sel.
Areas Commun., vol. 39, no. 4, pp. 903–918, Apr. 2021.[32] P. Bello, “Characterization of randomly time-variant linear channels,”
IEEE Trans. Commun. Syst., vol. 11, no. 4, pp. 360–393, Dec. 1963.[33] W. C. Jakes and D. C. Cox, Microwave Mobile Communications. New
York: IEEE Press, 1994.[34] S. Jaeckel, L. Raschkowski, K. Borner, and L. Thiele, “QuaDRiGa: A 3-
D multi-cell channel model with time evolution for enabling virtual fieldtrials,” IEEE Trans. Antennas Propag., vol. 62, no. 6, pp. 3242–3256,Jun. 2014.
[35] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.Cambridge university press, 2005.
[36] S. Jaeckel, L. Raschkowski, and L. Thiele, Quasi Deterministic Radio
Channel Generator User Manual and Documentation, Version 2.0.0,Fraunhofer Heinrich Hertz Institut, Aug. 2017. [Online]. Available:http://www.quadriga-channel-model.de/
[37] F. Hlawatsch and G. Matz, Wireless Communications Over Rapidly
Time-varying Channels. Academic press, 2011.
[38] C. Sun, X. Gao, S. Jin, M. Matthaiou, Z. Ding, and C. Xiao, “Beam divi-sion multiple access transmission for massive MIMO communications,”IEEE Trans. Commun., vol. 63, no. 6, pp. 2170–2184, Jun. 2015.
[39] W. U. Bajwa, J. Haupt, A. M. Sayeed, and R. Nowak, “Compressedchannel sensing: A new approach to estimating sparse multipath chan-nels,” Proc. IEEE, vol. 98, no. 6, pp. 1058–1076, Jun. 2010.
[40] E. J. Candes and T. Tao, “Decoding by linear programming,” IEEE
Trans. Inf. Theory, vol. 51, no. 12, pp. 4203–4215, Dec. 2005.[41] E. J. Candes and M. B. Wakin, “An introduction to compressive
sampling,” IEEE Signal Process. Mag., vol. 25, no. 2, pp. 21–30, Mar.2008.
[42] D. C. Chu, “Polyphase codes with good periodic correlation properties,”IEEE Trans. Inf. Theory, vol. 18, no. 4, pp. 531–532, Jul. 1972.
[43] L. R. Welch, “Lower bounds on the maximum cross correlation ofsignals,” IEEE Trans. Inf. Theory, vol. 20, no. 3, pp. 397–399, May1974.
[44] K. Wu, W. Wang, L. Gan, and X. Gao, “Compressive CSI acquisitionand non-orthogonal pilot design for downlink massive MIMO systems,”in Proc. IEEE Global Commun. Conf., Washington, DC, USA, Dec.2016, pp. 1–6.
Recommended